omnius 1.0.194 → 1.0.195

package/dist/index.js

@@ -552139,6 +552139,14 @@ function normalizeProviderToolMessage(msg, model) {
     toolCalls: toolCalls.length > 0 ? toolCalls : void 0
   };
 }
+function isGenerationToolName(toolName) {
+  return /^(generate_image|generate_audio|generate_video|generate_model|generate_tts|create_audio_file)$/.test(toolName);
+}
+function isGenerationArtifactSuccess(toolName, output) {
+  if (!isGenerationToolName(toolName))
+    return false;
+  return /(?:Image generated|Music generated|Sound generated|Video generated|3D model generated|CAD generated|Model generated|TTS generated|Created [A-Z]+ file|Created|Saved to|Output saved to):?\s+/i.test(output);
+}
 function inferEpisodeModality(toolName) {
   if (VISUAL_TOOLS.has(toolName))
     return "visual";
@@ -553638,6 +553646,8 @@ ${result.output ?? ""}`;
        * no todo list, returns an empty array to allow task completion.
        */
       getOpenTodoItems() {
+        if (this.options.disableTodoCompletionGuard)
+          return [];
         const todos = this.readSessionTodos();
         if (!todos || todos.length === 0)
           return [];
@@ -557803,7 +557813,7 @@ ${_staleSamples.join("\n")}` : ``,
             const userMsg = this.pendingUserMessages.shift();
             await this.appendInjectedUserMessage(userMsg, messages2, turn);
           }
-          {
+          if (!this.options.disableTodoPlanningNudges) {
             const maybeReminder = this.getTodoReminderContent(turn);
             if (maybeReminder) {
               messages2.push({ role: "user", content: maybeReminder });
@@ -557815,7 +557825,7 @@ ${_staleSamples.join("\n")}` : ``,
             }
           }
           const turnTier = this.options.modelTier ?? "large";
-          if (turn === 0 && (turnTier === "small" || turnTier === "medium")) {
+          if (turn === 0 && !this.options.disableTodoPlanningNudges && (turnTier === "small" || turnTier === "medium")) {
             const goal = this._taskState.goal || "";
             const wordCount2 = goal.split(/\s+/).length;
             const hasMultipleActions = /\band\b.*\band\b|then.*then|also.*also/i.test(goal);
@@ -558927,6 +558937,9 @@ ${criticDecision.cachedResult.slice(0, 500)}` : `[BLOCKED — the observer confi
                   mode: "step_repetition",
                   rationale: `force_progress_block on ${tc.name} after ${criticDecision.hitNumber} identical calls`
                 });
+                const generationCompletionHint = isGenerationArtifactSuccess(tc.name, criticDecision.cachedResult) ? `
+
+[GENERATION ALREADY COMPLETE] This exact ${tc.name} call already succeeded. Do not call it again. Use the cached artifact/path above; if delivery is needed, send it, otherwise call task_complete.` : "";
                 const header = criticDecision.compacted ? `[RE-SERVED FROM CACHE — the original result was compacted from context. Here is the data again. Do not retry this exact call.]
 
 ` : `[SKIPPED DUPLICATE — exact ${tc.name} call not re-run. The cached result below is from the prior successful call. Do not retry this exact call.]
@@ -558941,7 +558954,7 @@ ${truncatedCache}`);
                   tc,
                   output: `${criticDecision.blockMessage}
 
-${header}${truncatedCache}`,
+${header}${truncatedCache}${generationCompletionHint}`,
                   success: true
                 };
               }
@@ -558959,6 +558972,9 @@ ${header}${truncatedCache}`,
                   turn,
                   timestamp: (/* @__PURE__ */ new Date()).toISOString()
                 });
+                const generationCompletionHint = isGenerationArtifactSuccess(tc.name, criticDecision.cachedResult) ? `
+
+[GENERATION ALREADY COMPLETE] This exact ${tc.name} call already succeeded. Do not call it again. Use the cached artifact/path above; if delivery is needed, send it, otherwise call task_complete.` : "";
                 const header = criticDecision.compacted ? `[RE-SERVED FROM CACHE — the original result was compacted from context. Here is the data again. No need to call this tool again.]
 
 ` : `[DUPLICATE CALL #${criticDecision.hitNumber} — you already called ${tc.name} with these exact arguments. The result is identical. Do NOT call this again. Use the data you already have to make progress. One more identical call will trigger a hard progress block.]
@@ -558966,7 +558982,7 @@ ${header}${truncatedCache}`,
 `;
                 const truncatedCache = criticDecision.cachedResult.length > 500 ? criticDecision.cachedResult.slice(0, 500) + `
 ... [${criticDecision.cachedResult.length - 500} chars omitted — same as before]` : criticDecision.cachedResult;
-                const dedupOutput = header + truncatedCache;
+                const dedupOutput = header + truncatedCache + generationCompletionHint;
                 markSyntheticToolLog(dedupOutput);
                 this.emit({
                   type: "tool_result",
@@ -560160,6 +560176,9 @@ Do NOT retry ${tc.name} with similar arguments.`);
               } else if (result.success && tc.name !== "task_complete") {
                 sameToolFailStreak = 0;
                 sameToolFailName = null;
+                if (isGenerationArtifactSuccess(tc.name, result.output ?? "")) {
+                  this.pendingUserMessages.push(`[GENERATION COMPLETE] ${tc.name} succeeded. Do not call the same generation tool again for the same request. Use the artifact/path from the tool result; if delivery is needed, send it, otherwise call task_complete.`);
+                }
               }
               if (filePath && (tc.name === "file_read" || tc.name === "file_write" || tc.name === "file_edit" || tc.name === "batch_edit" || tc.name === "file_patch")) {
                 const isModify = tc.name !== "file_read";
@@ -634226,6 +634245,7 @@ Telegram response contract:
 - Do not summarize the fact that you answered; the visible assistant text must be the answer itself.
 - If you delegated long-running work, include the sub-agent id/status and what the admin should expect next.
 - Do not narrate retry strategy ("let me try one more approach", "actually", "wait, let me try") in assistant text. If you are stuck, send a single concise blocker sentence as the reply instead of streaming deliberation.
+- Treat todo_write as optional scratch state in Telegram runs. Do not create or maintain todos for simple Q&A, capability checks, or single artifact-generation requests.
 `.trim();
     TELEGRAM_EXTERNAL_ACQUISITION_CONTRACT = `
 External acquisition contract:
@@ -634412,10 +634432,6 @@ Telegram link integrity contract:
     TELEGRAM_ALLOWED_UPDATES = ["message", "guest_message", "callback_query", "poll", "message_reaction", "message_reaction_count"];
     TELEGRAM_DEFAULT_LONG_POLL_TIMEOUT_SECONDS = 50;
     TELEGRAM_DEFAULT_ROUTER_MODEL_CANDIDATES = [
-      "qwen3:0.6b",
-      "qwen3:1.7b",
-      "qwen3:4b",
-      "qwen3:8b",
       "qwen2.5:3b",
       "qwen2.5:7b",
       "llama3.2:1b",
@@ -634423,7 +634439,11 @@ Telegram link integrity contract:
       "gemma3:1b",
       "gemma3:4b",
       "phi3:mini",
-      "phi4-mini:latest"
+      "phi4-mini:latest",
+      "qwen3:0.6b",
+      "qwen3:1.7b",
+      "qwen3:4b",
+      "qwen3:8b"
     ];
     TELEGRAM_PUBLIC_TOOL_QUOTAS = {
       web: { limit: 20, windowMs: 60 * 6e4 },
@@ -639374,6 +639394,19 @@ ${retryText}`,
         const candidates = raw ? raw.split(/[,\s]+/).map((part) => part.trim()).filter(Boolean) : TELEGRAM_DEFAULT_ROUTER_MODEL_CANDIDATES;
         return Array.from(new Set(candidates));
       }
+      telegramRouterAllowThinkHeavyAutoModels() {
+        const raw = (process.env["OMNIUS_TG_ROUTER_ALLOW_THINK_MODELS"] ?? "").trim().toLowerCase();
+        return raw === "1" || raw === "true" || raw === "on";
+      }
+      telegramRouterModelLooksThinkHeavy(name10) {
+        return /\b(?:qwen3|qwq|deepseek-r1|r1-|reasoning)\b/i.test(name10);
+      }
+      orderTelegramRouterCandidates(candidates) {
+        if (this.telegramRouterAllowThinkHeavyAutoModels()) return candidates;
+        const stable = candidates.filter((candidate) => !this.telegramRouterModelLooksThinkHeavy(candidate));
+        const thinkHeavy = candidates.filter((candidate) => this.telegramRouterModelLooksThinkHeavy(candidate));
+        return [...stable, ...thinkHeavy];
+      }
       normalizeOllamaModelNameForMatch(name10) {
         return name10.trim().toLowerCase().replace(/:latest$/, "");
       }
@@ -639404,7 +639437,7 @@ ${retryText}`,
             source: "main"
           };
         }
-        const candidates = this.telegramRouterCandidateModels();
+        const candidates = this.orderTelegramRouterCandidates(this.telegramRouterCandidateModels());
         const cacheKey = `${config.backendUrl}
 ${config.model}
 ${candidates.join(",")}`;
@@ -639432,7 +639465,7 @@ ${candidates.join(",")}`;
               atMs: now,
               model: selected,
               source: "auto-small",
-              detail: "selected first installed OMNIUS_TG_ROUTER_MODEL_CANDIDATES entry from Ollama /api/tags"
+              detail: "selected first installed Telegram router candidate from Ollama /api/tags; think-heavy models are tried last unless OMNIUS_TG_ROUTER_ALLOW_THINK_MODELS=1"
             };
             this.telegramRouterModelCache = resolved;
             return {
@@ -640594,7 +640627,7 @@ ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
             return result.result?.message_id ?? null;
           } catch (err) {
             const errStr = err instanceof Error ? err.message : String(err);
-            if (this.shouldLogTelegramSendFailure(errStr)) {
+            if (process.env["OMNIUS_TELEGRAM_DEBUG_SEND_FAILURES"] === "1" && this.shouldLogTelegramSendFailure(errStr)) {
               this.tuiWrite(() => renderWarning(`Failed to send Telegram live message: ${errStr}`));
             }
             this.updateTelegramTextDeliveryCapability(chatId, {
@@ -642020,6 +642053,8 @@ ${conversationStream}`
           compactionThreshold: this.telegramFallbackCompactionThreshold(modelTier),
           contextWindowSize,
           modelTier,
+          disableTodoCompletionGuard: true,
+          disableTodoPlanningNudges: true,
           streamEnabled: true,
           dynamicContext: sessionContext.context,
           captureContextFrame: true,

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.194",
+  "version": "1.0.195",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.194",
+      "version": "1.0.195",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.194",
+  "version": "1.0.195",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",